The reverse mathematics of Hindman's Theorem
Paul-Elliot Anglès d'Auriac
Joint work with Ludovic Patey and Benoît Monin
Hindman's Theorem
Let \(A\subseteq\mathbb N\). Then, we define the set of sums of finitely many elements from \(A\). $$\operatorname{FS}(A) = \left\{\sum_{i=0}^na_i\quad:\quad n\in\mathbb N,\ \ a_i\in A,\ \ \text{ the \(a_i\) are all different}\right\}$$
Note that \(\mathrm{FS}(A)\) is not closed under finite sums. For instance, \[\operatorname{FS}(\{1\}\cup3\mathbb N)=3\mathbb N\cup(3\mathbb N+1)\]
For all colorings of the integers in a finite number of colors \(\mathbb N=C_0\cup\cdots\cup C_{n-1}\), there exists a color \(i\) and an infinite set \(H\subseteq \mathbb N\) and such that \[\operatorname{FS}(H)\subseteq C_i\]
  • The coloring of even and odd integers,
  • The \(\mathrm{mod}\ 3\) coloring,
  • The coloring given by the parity of the length of the binary expansion
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
...

A more general formulation

An IP set is a set \(A\subseteq\mathbb N\) such that there exists an infinite \(S\) such that \(\operatorname{FS}(S)\subseteq A\)
  • \(\mathbb N\) is an IP set, but no finite set in an IP set,
  • If \(A\) is an IP set and \(B\supseteq A\), then \(B\) is an IP set.
A class \(\mathcal C\subseteq\mathcal P(\mathbb N)\) is partition regular, if for every \(A\in\mathcal C\), if \(A\) is partitionned in \(A=A_0\cup\dots\cup A_{n-1}\), then for some \(i\lt n\), \(A_i\in\mathcal C\).
The IP sets form a partition regular class. For any IP set \(A=A_0\cup\cdots\cup A_{n-1}\), for some \(i\lt n\), \(A_i\) is an IP set.
  • This reveals the "Ramseyan" statement: for every coloring of a structure, there is monochromatic substructure
  • This allows to iterate Hindman's Theorem!

State of the art of the Reverse Mathematics of Hindman's theorem.

For every finite coloring \(c:\mathbb N\to2\), there exists a monochromatic IP set \(H\) computable in the \((\omega+1)\)-th Turing jump of \(c\).
Hindman Theorem is provable in \(\mathrm{ACA}_0^+\).
There exists a computable finite coloring such that all homogeneous sets \(H\) compute \(\emptyset'\), and one with no \(\Sigma^0_2\) solution.
Hindman's theorem proves \(\mathrm{ACA}_0\).
Given a computable coloring \(c\):
  • \(\emptyset^{(\omega+1)}\) computes a monochromatic IP set.
  • \(\emptyset^{'}\) might not compute a monochromatic IP set.
What is the exact \(\alpha\) such that:
  • \(\emptyset^{(\alpha)}\) computes a monochromatic IP set,
  • \(\emptyset^{(\beta)}\) might not compute a monochromatic IP set for \(\beta<\alpha\)?
In terms of reverse mathematics:
  • Is Hindman Theorem provable in \(\mathrm{ACA}_0\)?
    (~corresponds to \(\alpha\lt\omega\))
  • Does Hindman Theorem proves in \(\mathrm{ACA}^+_0\)?
    (~corresponds to \(\alpha\geq\omega\))
  • Is it strictly in between?
    (due to lack of induction)
Effective proof of Hindman Theorem
Define a largeness notion such that we can prove: $$\text{The large sets form a partition regular class}\tag{1}$$ $$A\text{ is large }\Longrightarrow \exists n\in A, A\cap A+n\text{ is large. }\tag{2}$$
What are the possible candidate for this notion of largeness? In particular, a large set must be an IP set!
  • (2) does not hold for large being infinite
  • (1) is equivalent to Hindman's Theorem when large is being an IP set

Example of largeness notions that work

Define a largeness notion such that we can prove: $$\text{The large sets form a partition regular class}\tag{1}$$ $$A\text{ is large }\Longrightarrow \exists n\in A, A\cap A+n\text{ is large. }\tag{2}$$
An idempotent ultrafilter is an ultrafilter \(\mathcal U\) such that \(\mathcal U = \mathcal U + \mathcal U\), where \[A\in\mathcal U + \mathcal U \Longleftrightarrow \{n:A-n\in\mathcal U\}\in\mathcal U\]
For any idempotent ultrafilter \(\mathcal U\),
  • If \(A=A_0\cup\dots\cup A_{n-1}\in\mathcal U\), then \(\exists i\lt n\) s.t. \(A_i\in\mathcal U\),
  • If \(A\in\mathcal U\), then \(\exists n\in A\) s.t. \(A\cap A+n\in\mathcal U\).
If \(I\) is an IP set, a set \(A\) is said to be large in \(I\) if there is no IP set \(I'\subseteq I\), such that \(I'\cap A=\emptyset\). It is B-large if it is large in some IP set \(I\).
If \(A\) is large in \(I\), there is no way to avoid color \(A\) by shrinking \(I\)!
  • If \(A=\bigcup_{i\lt n}A_i\) is B-large, then \(\exists i\lt n\) s.t. \(A_i\) is B-large,
  • If \(A\) is B-large, then \(\exists n\in A\) s.t. \(A\cap A+n\) is B-large.
As a consequence, a set is B-large if and only if it is an IP set.

Effectivizing the proof

Recall
$$A\text{ is large }\Longrightarrow \exists n\in A, A\cap (A+n)\text{ is large. }\tag{2}$$
The problem is that the \(n\) found by (2) is chosen in a non uniform way, out of infinitely many, and iterated \(\omega\) many times. Consider:
$$A\text{ is large }\Longrightarrow \exists {\color{maroon}n_0,\cdots, n_N}\in A, \bigcup_{i\lt N}A\cap (A+{\color{maroon}n_i})\text{ is large. }\tag{2'}$$
We have that \((1)+(2')\Rightarrow(2)\):
$$A\text{ is large }\xRightarrow{(2')}\bigcup_{n\in F}A\cap (A+n)\text{ is large }\xRightarrow{(1)}\exists n\in F, A\cap (A+n)\text{ is large.}$$
This time, \(n\) is found by (1) in a non uniform way, but out of finitely many. We postpone this choice to the end, and split the construction.

Avoiding the use of (1) in the proof, supposing (2')

$$A\text{ is large }\Longrightarrow \exists {\color{maroon}n_0,\cdots, n_N}\in A, \bigcup_{i\lt N}A\cap (A+{\color{maroon}n_i})\text{ is large. }$$
A full-match for a coloring \(c\) is a couple \(({\color{maroon}F},S)\) where \(F\) is a finite set, \(S\) is an IP set, and for every \(x\in S\), there exists \({\color{maroon}a}\in {\color{maroon}F}\) such that \[c({\color{maroon}a})=c(x)=c({\color{maroon}a}+x)\]
The complexity of a full-match is essential to understand the complexity of a solution to Hindman's theorem!

Complexity of full/right/left-matches

A full-match for a coloring \(c\) is a couple \(({\color{maroon}F},S)\) where \(F\) is a finite set, \(S\) is an IP set, and for every \(x\in S\), there exists \({\color{maroon}a}\in {\color{maroon}F}\) such that \[c({\color{maroon}a})=c(x)=c({\color{maroon}a}+x)\]
Does there always exists a computable full-match given a computable coloring?
  • An affirmative answer would prove Hindman's theorem equivalent to \(\mathrm{ACA}_0\)
  • A negative answer to the question was published, but there was an error in the proof.
The proof of the existence of a full-match is similar to Hindman's Theorem, given the existence of right-matches. The proof yields an arithmetical solution.
A right-match for a coloring \(c\) is a couple \((F,S)\) where \(F\) is a finite set, and for every \(x\in\operatorname{FS}(S)\), there exists \(a\in F\) such that \[c(x)=c(a+x)\]
Every computable coloring admits a computable right-match, uniformly in the double jump of the coloring.
A left-match for a coloring \(c\) is a couple \((F,S)\) where \(F\) is a finite set, and for every \(x\in\operatorname{FS}(S)\), there exists \(a\in F\) such that \[c(a)=c(a+x)\]
Every computable coloring in two colors admits a computable left-match.
However, the following hold:
  • The existence of computable full-match for 2 colors implies the existence of computable full-matches for \(n\) colors, but
  • The existence of computable left-match for 2 colors does not imply the existence of computable left-matches for \(n\) colors.
Does there always exist a computable full-match given a computable 2-coloring?
Does there always exist a computable left-match, given a computable 3-coloring?
Consider the following property: $$\exists F\subseteq C_i, \forall x\in\operatorname{FS}(S),\ \exists a\in F,\ c(a+x) = j\tag{\(\mathrm{MM}_{i,j}\)}$$
Every coloring in two colors satisfying one of the following properties:
  • \(\lnot(\mathrm{MM}_{0,1})\) or \(\lnot(\mathrm{MM}_{1,0})\)
  • \((\mathrm{MM}_{0,0})\) and \((\mathrm{MM}_{1,1})\)
  • \(\lnot(\mathrm{MM}_{0,0})\) and \(\lnot(\mathrm{MM}_{1,1})\)
admits a computable full-match.
Thank you for your attention!